Mistral launches Voxtral 3 speech models and a voice agent SDK, with Hindi and six other Indian languages

What Mistral launched

Voxtral 3 is a set of speech models plus a toolkit for building voice agents, available through Mistral’s EU, US and global API endpoints or on a company’s own infrastructure.¹²

• Voxtral Small Transcribe 3: Mistral’s most accurate speech-to-text model, with speaker labelling, keyword hints and end-of-turn detection.¹²
• Voxtral Mini Transcribe 3: a 4-billion-parameter speech-to-text model for batch work.¹²
• Voxtral Mini Transcribe Realtime 3: streaming speech-to-text for live agents.¹
• Voxtral Mini TTS 3: text-to-speech with control over emotion and pacing, and voice cloning.¹²

Languages, including Indian ones

Transcription covers more than 50 languages, and realtime transcription more than 35.¹² Mistral lists Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi and Telugu for transcription.¹ Text-to-speech covers more than 35 languages and 50 dialects; Hindi is named, other Indian languages are not listed for speech output.¹

Prices

| Product | Price | |---|---| | Mini Transcribe 3 (batch) | $0.0035 per minute | | Small Transcribe 3 | $0.004 per minute | | Mini Transcribe Realtime 3 | $0.008 per minute | | Mini TTS 3 | $13 per million characters | | Voice agent with Mistral Small 4 | $0.04 per minute |

Prices are from Mistral’s launch page.¹ TestingCatalog reports the same transcription and speech prices.²

The voice agent SDK

The SDK is built on LiveKit and runs listening, reasoning, tool use and speaking together, keeping turns under a second and stopping when a caller interrupts, according to TestingCatalog.² It connects by phone (SIP) or web (WebRTC), and each part can be swapped.¹² Self-hosting keeps audio inside the customer’s own environment.²

Accuracy

Mistral says Small Transcribe 3 leads on the 45-language FLEURS benchmark, and TestingCatalog reports about a 4% error rate on Mistral’s chart.¹² These are Mistral’s own comparisons. Mistral did not say whether any Voxtral 3 weights are open.¹²

What it means

For Indian businesses, the language list is the headline. Call centres, banks, clinics and government services that handle calls in Hindi, Marathi or Telugu can now test a major provider’s transcription in those languages at low per-minute prices. The ability to self-host also helps where audio must stay inside a company’s own systems. The gap to check is speech output: only Hindi is named for text-to-speech, so a fully spoken agent in other Indian languages may still need another provider.

What to watch

• Independent accuracy tests on Indian-language speech, including mixed Hindi-English.
• Whether Mistral adds more Indian languages to text-to-speech.

Related: Mistral and Nvidia’s engineering AI plans and ElevenLabs’ Eleven Realtime voice models.